摘要
针对现有网络信息推荐方法普遍存在关注主题局限,潜在兴趣挖掘不准确等问题,提出了融合LDA与注意力的网络信息个性化推荐算法。首先采用LDA模型归纳文档中主题与词的分布情况,为保证兴趣单词选取的完整性,引入HowNet来处理单词语义,以义原为最小单位来描述,根据分层机制逐层推导至单词相似性,从而避免通过距离来计算语义间的相似性。然后采取注意力机制为关注事务服务,针对特征词的重要性,利用加权值的改变,优化推荐内容的精确程度。将注意力注入到网络层中,在注意力层中,先启动实体部分,完成重要语义实体注入;再启动语义部分,根据传输路径的关注度,注入合理的相近实体。最后,基于Amazon开放的TH和SO数据集对推荐算法进行仿真分析,采取MSE、HR和NDCG三个指标进行衡量。通过实验结果,证明所提方法在推荐精度和误差方面均获得了有效改善,且对于不同数据集具有良好的适用性和泛化能力,推荐结果高度符合用户的真实需求。
Aiming at the problems of the existing network information recommendation methods, such as the limitation of focusing on topics and inaccurate mining of potential interests, a personalized recommendation algorithm for network information is proposed, which combines attention with fusion. Firstly, LDA model was used to summarize the distribution of topics and words in the document. In order to ensure the integrity of the selection of words of interest, HowNet was introduced to process the word semantics, which was described with the original meaning as the smallest unit. According to the hierarchical mechanism, the word similarity was deduced layer by layer, so as to avoid calculating the semantic similarity through distance. Then the attention mechanism was adopted to serve the concerned affairs. According to the importance of feature words, the accuracy of recommended content was optimized by changing the weighted value. The attention was injected into the network layer;Next, in the attention layer, the entity part was started first to complete the injection of important semantic entities;Then the semantic part was started and the reasonable similar entities were injected according to the attention of the transmission path. Finally, the recommended algorithm was simulated and analyzed based on Amazon’s open TH and SO data sets, and measured by MSE,HR and NDCG. The experimental results show that the proposed method has effectively improved the recommendation accuracy and error, and has good applicability and generalization ability for different data sets. The recommendation results highly meet the real needs of users.
作者
张永宾
赵金楼
ZHANG Yong-bin;ZHAO Jin-lou(School of Economics and Management,Harbin Engineening University,Harbin Heilongjiang 150001,China;School of Management,Heilongjiang University of Science and Technology,Heilongjiang 150022,China)
出处
《计算机仿真》
北大核心
2022年第12期528-532,共5页
Computer Simulation
基金
国家自然基金(71540019)
教育部人文社会科学青年基金项目(20YJCZH035)。
关键词
注意力机制
推荐算法
主题分布
Attention mechanism
Recommendation algorithm
Subject distribution
